Two closely named NAVER services solve different transcription problems. CLOVA Speech targets longer recordings and subtitle workflows; CLOVA Speech Recognition, often called CSR, was designed for short command-like utterances.
The two products are not interchangeable
NAVER’s API documentation distinguishes CLOVA Speech for long audio and video uploads from the older CSR service for short imperative speech under about a minute. The names are similar enough that a global product directory can accidentally merge their endpoints, latency targets, billing paths and authentication schemes. Maintain separate product records and never assume that a feature from one is supported by the other without an explicit API contract.
Speech for recorded media
CLOVA Speech provides APIs supporting transcription of long audio/video, call transcripts and subtitles. It uses a domain-specific API Gateway invoke URL, which is assigned per account and deployment. That means a sample address cannot be copied directly into an independent application. Keep both the object-storage input policy and generated transcript retention policy in view; long recordings can include multiple speakers and private information.
CSR for shorter voice input
CSR exposes RESTful speech-to-text and SDK-oriented mobile workflows. The documented REST request uses the /stt path and a language parameter, with examples for Korean, English, Japanese and Chinese. Acceptable input encodings include familiar compressed and lossless formats, but actual sample rates and transport constraints should be checked against the endpoint documentation rather than inferred from a file extension. A one-minute voice command use case should not be billed or benchmarked as a one-hour media transcript.
Multilingual and regional review
Korean recognition needs a representative Korean-language test corpus, and English or Japanese support alone does not establish dialect accuracy. Run evaluations with speaker age, environmental noise, overlapping dialogue and proper names taken from the intended audience. NAVER Cloud deployment access and support policies may vary across regional accounts. Record locale and account region in every quality result and avoid assuming a global self-service entitlement.
Implementation checklist
Create a separate environment config for each service: credentials, assigned invoke URL, audio ingress, duration validation, callbacks or polling, transcript format, privacy notices and quota handling. Build explicit routing rules using content length and expected interactive latency. Preserve a human review fallback for name-sensitive speech, subtitles and legal content; neither interface verifies a speaker’s legal identity.
Pricing, permissions and date scope
All changing statements are scoped to documentation consulted by 25 October 2026. Model and account support, rights, language coverage, endpoint configuration and regional access can change. This article does not assert first-hand testing, an independent benchmark, present-day tariff, guaranteed authorship or unconditional commercial rights. Recheck the exact subscription, license and supported region before production.